Trigger word: tgrss

Master Tigress from the Kung Fu Panda films, trained on Krea 2. Built to stay in character while leaving the rest of the image up to you — outfit, lighting, framing and style all remain promptable, and it stacks cleanly with realism and style LoRAs.

Settings

These are the exact settings the sample images were made with:

Model - Krea 2 Turbo

LoRA weight - 1.0 (0.8 if layering other LoRAs)

Steps - 8

CFG / guidance - 1.0 (CFG off — Turbo does not use a negative prompt)

Sampler - euler

Scheduler - simple

Denoise - 1.0

How to prompt it

The captions all followed one shape, so prompts built the same way land closest to what it knows:

tgrss, an anthropomorphic tiger in [outfit], [view / pose], [expression], [lighting], [background], [framing], film still

tgrss must lead. The phrase an anthropomorphic tiger appears in every single caption, so including it helps.

Framing words it actually knows, with how many of the 33 captions used each: bust framing (13) · medium shot (14) · full body (3) · face close-up (1)

The film still tail is a style switch. 31 of 33 captions end with it and the other 2 end with clean promotional render:

• keep film still for the movie look

• swap to clean promotional render for the white-background studio-render look

• drop it entirely when stacking a realism/photo LoRA — otherwise it fights the realism and drags the image back toward CG

What it does well

• Outfits are promptable. The red training vest, the yellow/gold tunic, or something the character has never worn — a winter coat, modern clothes, whatever you ask for.

• Angles it was never trained on — strict profile, overhead, low angle — generalise fine.

• Close-ups hold up. The dataset was cut from 4K sources, so fur and eye detail survive at portrait framing rather than going soft.

• Composes rather than dominates. This targets attention layers only, leaving the MLP layers untouched, so it steers identity without overriding the style of whatever else you have loaded.

Example prompts

1 — portrait, film look (the baseline)

tgrss, an anthropomorphic tiger in a dark red sleeveless vest with gold leaf embroidery, head and shoulders portrait, three-quarter view, calm confident expression with slightly narrowed amber eyes, soft warm studio lighting from the left, plain dark grey background, bust framing, film still

2 — full scene

tgrss, an anthropomorphic tiger in a dark red sleeveless vest with gold leaf embroidery and charcoal trousers, three-quarter view walking through a village street with a faint confident smile, white walls and tiled roofs behind, bright overcast daylight, casual stride, medium shot, film still

3 — an outfit she has never worn

tgrss, an anthropomorphic tiger wearing a plain blue winter coat, standing in falling snow at dusk, calm neutral expression, cool blue light, medium shot, film still 

4 — stacked with a realism LoRA (drop film still)

tgrss, an anthropomorphic tiger in a dark red sleeveless vest, candid photo, shallow depth of field, natural window light, bust framing

Tested with lenovo_ultrareal_krea2 chained after this LoRA. Identity stays intact across the whole useful range — tigress 0.8–1.0 against realism 0.6–1.0 — so you have room to balance to taste. A good starting point is tigress 0.9 / realism 0.8.

Note this shifts lighting, depth of field and framing toward photographic; it does not turn her photoreal, and dialling the realism LoRA higher does not force it. Removing film still from the prompt matters here — leaving it in pulls the render back toward the movie look and cancels most of what the realism LoRA is doing.

5 — photoreal / "real world" candid (the showcase look)

Three things make this work, and the prompt does as much of the job as the LoRA stack: push her into the real world, ask explicitly for fur rendering cues, and add a block of camera artefacts.

tgrss, found footage, a realistic tiger woman with anthropomorphic features in the real world, dense individual fur strands visible, subsurface scattering on ears and muzzle, muscular athletic build, broad shoulders, calm confident expression with a faint smirk, standing on a city sidewalk at night, hands in pockets, looking back over her shoulder at the camera, worn dark bomber jacket, wet pavement, colour cast from sodium streetlights, early 2010s phone photo quality, slight motion blur, harsh on-camera flash, background falls off quickly, digital noise in shadows, casual candid snapshot

LoRA chain used (order matters, all model-only):

Krea2-realism-V2 - 0.30

lenovo_ultrareal_krea2 - 1.20

nicegirls_krea2 - 0.40

tigress_krea2_v1 - 1.05

Note film still is absent — including it cancels most of the realism. Expect her head to read slightly large against the body at these weights; that is the trade for keeping identity strong, and lowering the character weight to 0.9 barely changes it.

Known limitations

• The red vest is over-represented in the training data (roughly 3:1 against the yellow tunic). Both work, but at high weight combined with a strong competing style LoRA, red may try to reassert itself. Drop to 0.8 if that happens.

• Trained on the film character design, not the TV series look.

• Adult design only — no cub/young versions.

Training

Base - Krea 2 RAW (inference on Turbo)

Network - LoRA, rank 64 / alpha 64, attention layers only

Dataset - 33 images, hand-captioned

Source - 4K UHD film frames, HDR tone-mapped, plus a few studio promo renders

Steps - 2,970 (90 epochs), lr 1e-4, adamw8bit

Trainer - musubi-tuner

Dataset was curated by hand rather than scraped — every frame checked individually for clipped framing, motion blur, wrong character and duplicates, and every caption written to name the outfit explicitly so the trigger carries identity rather than clothing